Most AI jobs end at a prediction. Here the prediction becomes a setpoint on a 2,000 litre vessel at 3am, and a chemist has to trust it. If that sounds like the interesting version of the problem, we should talk.
All roles involve time inside real plants. Nobody here builds manufacturing autonomy purely from a laptop.
ENGINEERING
PROCESS ML ENGINEER
Build the process world model: multimodal models over inline sensors, formula structure and batch outcomes. You should be comfortable being wrong in front of a chemist.
PyTorch, time series, multimodal
Bonus: rheology or chemical engineering
ENGINEERING
EDGE SYSTEMS ENGINEER
Deterministic Rust/C++ runtime on Jetson and IGX: control loops, watchdogs, safe-state, offline survival and signed OTA.
Rust or C++, Linux, real-time
Bonus: OPC-UA, industrial protocols
ENGINEERING
COMPUTER VISION ENGINEER
Spectral shade regression and high-speed defect vision across compacts, bullets, jars and tubes at full line rate.
TensorRT, camera pipelines
Bonus: colour science, spectrophotometry
ENGINEERING
OPTIMISATION ENGINEER
Campaign sequencing, changeover, cleaning validation and filler balancing with cuOpt at plant scale.
Constrained optimisation, OR
Bonus: manufacturing scheduling
FIELD
FORWARD-DEPLOYED ENGINEER
Live on site during instrumentation and go-live. Translate between chemists, OT teams and the model stack.
Strong generalist engineer
Bonus: GMP plant experience
QUALITY
QUALITY SYSTEMS LEAD
Own validation approach, audit design and ISO 22716 alignment. Make autonomy defensible to an inspector.
Cosmetics or pharma GMP
Bonus: computerised system validation
HOW WE WORK
The operating style
ON SITE, OFTEN
Engineers visit plants. You cannot model an emulsion you have never watched being made, and you cannot design a console for an operator you have never met.
HONEST ABOUT LIMITS
We label what is aspirational, internally and publicly. Overclaiming to a quality manager is a career-limiting move in this industry.
SMALL SURFACE, DEEP WORK
One vertical, five products, one runtime. We say no to adjacent verticals constantly.
WRITTEN THINKING
Decisions are written down with their assumptions, because in two years an auditor or a successor will ask why.
REAL DEADLINES
A pilot window is a plant’s production calendar. It does not move because a sprint slipped.
NO THEATRE
No demo-driven development. If it will not survive a night shift, it is not done.
HIRING PROCESS
Four steps, two weeks
01
INTRO
Thirty minutes on what you have built and what you want to build. Mutual filter, no whiteboard.
→
02
TECHNICAL
A real problem from our domain, discussed together. No leetcode, no take-home that eats a weekend.
→
03
DEPTH
A session with the team you would join, plus a conversation about how you handle being wrong in production.
→
04
OFFER
Reference calls, offer, and a written explanation of the level and the compensation logic.
WHAT WE OFFER
The deal
EQUITY
Meaningful equity at an early stage, explained clearly including the parts that are uncomfortable.
FLEXIBILITY
Hybrid working around plant visits. Deep work is protected; site weeks are intense by design.
HARDWARE
Real GPU access for development and training, plus the edge hardware you need on your desk.
LEARNING
You will learn cosmetics chemistry, GMP and industrial control from people who do it professionally.
LEVELS
How we think about seniority
How we think about seniority
LEVEL
SCOPE
SIGNAL
Engineer
Owns components
Ships correct work with support; asks good questions early
Senior
Owns a subsystem
Designs for failure modes; unblocks others; writes it down
Staff
Owns a domain
Sets technical direction; trusted alone in front of a customer’s quality team
Principal
Owns a cross-cutting bet
Changes what the company builds; carries the hardest ambiguity
REALITY CHECK
What this job is not
NOTA PURE RESEARCH ROLE
NOTFULLY REMOTE
NOTLOW STAKES
NOTA BIG TEAM
INTERVIEW PREP
What we will actually ask you
We ask about a system you shipped that had consequences when it failed, and what you changed afterwards. We ask how you would detect that a model has quietly degraded on one vessel.
We ask what you would refuse to automate. Candidates who cannot name anything they would refuse to automate tend not to do well here.
A failure you owned and what it changed.
How you would detect silent model degradation.
What you would refuse to automate, and why.
How you explain uncertainty to a non-ML expert.
The last time you were wrong in front of a customer.
PROCESS FACTS
STEPS
4
TYPICAL DURATION
2 WEEKS
TAKE-HOME
NO
LEETCODE
NO
SITE VISIT
OFFERED
FEEDBACK
ALWAYS GIVEN
“
If it will not survive a night shift with a tired operator and a sensor that just went noisy, it is not finished.
Serumon engineering standard
CAREERS FAQ
Before you apply
No. Roughly half the team will not have it. You do need genuine curiosity about the physical process, because the people you build for spent decades learning it.
Hybrid, organised around design-partner plant locations. Specific locations are discussed at the intro stage and shift as deployments land.
Not currently. The work requires plant access and production responsibility that is difficult to structure for a short placement [ASPIRATIONAL for later stages].
Send a speculative application describing what you would build here. Specific, opinionated proposals get read carefully; generic CVs do not.
APPLY
Tell us what you would build
Not a cover letter. One page on the hardest part of closing a control loop in a GMP plant, and how you would attack it.